علي جمالي زواره

Given the increasing use of augmented reality in people's daily lives, the need to develop more accurate and cost-effective methods in this field seems essential. In this research, after introducing the background of studies on the computation of the fundamental matrix, which is one of the foundational methods for achieving augmented reality algorithms, new hybrid methods are introduced by combining deep networks and traditional techniques. In these hybrid methods, each of the three key stages in the computation of the fundamental matrix is performed using one of the existing deep or traditional methods, and the fundamental matrix computed in the final stage is based on the results of the previous steps. In the end, the efficiency of these hybrid methods is eva‎luated and compared using two different approaches and two criteria: time and accuracy. The results from the eva‎luations show that the method which uses SIFT for keypoint detection, Brute-force for matching, and DFE for computing the fundamental matrix has higher accuracy compared to other methods in estimating the fundamental matrix. Additionally, in the 3D point projection test on the image, the method that uses SuperPoint for keypoint detection, SuperGlue for matching, and RANSAC for computing the fundamental matrix achieved higher accuracy than the other methods.

ماتريس بنيادي , واقعيت افزوده , شبكه عصبي , يادگيري عميق

Fundamental matrix , Augmented reality , Neural networks , Deep learning

Ali Jamali Zavareh

Mohsen Soryani